Self-supervised transaction embeddings improve fraud detection and risk analysis
CARTA: Self-Supervised Reusable Transaction Embeddings via Heterogeneous Tokenization and Cross-Attention
Machine LearningComputation and Language
Summary
Detecting fraud in payment transactions is hard because labels are expensive to get and some transaction details like merchant names vary a lot, making standard methods inefficient. The authors built CARTA, a model that learns a small set of numbers (an embedding) to describe transactions without needing any labels. This reusable representation works well for detecting fraud even with very few fraud labels and helps group transactions by risk without retraining. Their approach uses special handling of different data types and contrasts parts of transactions to learn meaningful patterns.
What this means in practice
- •For financial fraud teams: Use pre-trained transaction embeddings to detect fraud efficiently with minimal labeled examples in payment processing systems.
- •For risk analytics teams: Apply reusable transaction representations for unsupervised risk grouping of payments without needing to retrain models for each task.
Authors
Shreyas K. Chandrahas, Jose Munoz, Satish Babu Bellapu
Abstract
Acquiring fraud labels for payment transactions is expensive and inconsistent across tasks, and high-cardinality identifiers such as merchant names make one-hot feature pipelines infeasible at scale. We present CARTA, a self-supervised encoder that learns a compact 16-dim embedding from raw transactions without labels; the frozen embedding is then reused for label-efficient fraud probing and exploratory unsupervised risk stratification without retraining the encoder. CARTA uses type-specific projections with learned embedding tables for high-cardinality identifiers, trained with token excision and InfoNCE contrastive loss. On IEEE-CIS, its linear probe at 1% of labels exceeds a raw-feature probe at 100%; on IBM TabFormer (2M rows), its frozen embedding reaches GBDT AUPRC 0.508, above frequency encoding (0.411) and the merchant-dropping baseline (0.022), though fully-supervised CatBoost is stronger. CARTA is a label-free, reusable representation that complements task-specific supervised models.